Source-linked AI summary
Occluded Prohibited Items Detection: an X-ray Security Inspection Benchmark and De-occlusion Attention Module
Yanlu Wei, Renshuai Tao, Zhangjie Wu, Yuqing Ma, Libo Zhang, Xianglong Liu
TL;DR
Occluded overlaps in luggage make prohibited-item detection from X-ray images difficult, while existing studies and datasets provide limited coverage of this setting. The paper introduces the OPIXray benchmark and DOAM, which uses edge and material information to refine detector features. DOAM consistently improves popular detectors, with stronger advantages at higher occlusion levels, while its reported complexity increase is small.
Problem
Heavily overlapped luggage objects make accurate prohibited-item detection difficult, and few studies or datasets address occluded security-inspection X-ray detection.
Method
The paper builds the OPIXray cutter-detection benchmark and proposes DOAM, which combines edge guidance and material awareness for popular detectors.
Results
DOAM consistently improves state-of-the-art detectors and significantly outperforms several widely used attention mechanisms, with larger advantages at higher occlusion levels.
Takeaways & Limitations
OPIXray and DOAM support evaluation and detection of occluded prohibited items in realistic security-inspection X-ray settings.
Takeaways & Limitations
Different values of parameter k cause repetitive computation in RIA.
Abstract
from arXiv · showhide
Security inspection often deals with a piece of baggage or suitcase where objects are heavily overlapped with each other, resulting in an unsatisfactory performance for prohibited items detection in X-ray images. In the literature, there have been rare studies and datasets touching this important topic. In this work, we contribute the first high-quality object detection dataset for security inspection, named Occluded Prohibited Items X-ray (OPIXray) image benchmark. OPIXray focused on the widely-occurred prohibited item "cutter", annotated manually by professional inspectors from the international airport. The test set is further divided into three occlusion levels to better understand the performance of detectors. Furthermore, to deal with the occlusion in X-ray images detection, we propose the De-occlusion Attention Module (DOAM), a plug-and-play module that can be easily inserted into and thus promote most popular detectors. Despite the heavy occlusion in X-ray imaging, shape appearance of objects can be preserved well, and meanwhile different materials visually appear with different colors and textures. Motivated by these observations, our DOAM simultaneously leverages the different appearance information of the prohibited item to generate the attention map, which helps refine feature maps for the general detectors. We comprehensively evaluate our module on the OPIXray dataset, and demonstrate that our module can consistently improve the performance of the state-of-the-art detection methods such as SSD, FCOS, etc, and significantly outperforms several widely-used attention mechanisms. In particular, the advantages of DOAM are more significant in the scenarios with higher levels of occlusion, which demonstrates its potential application in real-world inspections. The OPIXray benchmark and our model are released at https://github.com/OPIXray-author/OPIXray.
1 INTRODUCTION
The paper addresses inter-class occlusion in X-ray security inspection by introducing the OPIXray benchmark and the DOAM module for occlusion-aware detection.
- Security inspection images contain heavily overlapped luggage objects, making accurate prohibited-item detection difficult for inspectors and automated systems.
- OPIXray is presented as the first high-quality object-detection benchmark specifically targeting occluded prohibited items in security-inspection X-ray images.It focuses on cutters and uses manual annotations by professional airport security inspectors.
- OPIXray contains 8,885 X-ray images covering five cutter categories, with manually localized prohibited items and realistic overlap patterns.
- The proposed De-occlusion Attention Module (DOAM) is a plug-and-play component that emphasizes edge and material information in occluded X-ray objects.Its Edge Guidance and Material Awareness sub-modules exploit preserved shape information and material-dependent colors and textures.
- DOAM can be inserted into popular detectors and consistently improves state-of-the-art detection methods while outperforming several widely used attention mechanisms.
2 RELATED WORK
Prior X-ray benchmarks and attention methods provide relevant foundations, but existing security-inspection datasets offer limited support for realistic occluded prohibited-item detection.
- 2.1 X-ray Images and Benchmarks: X-ray imaging is used in security inspection, but object occlusion substantially reduces the visibility of object information.
- 2.1 X-ray Images and Benchmarks: GDXray includes prohibited-item categories but uses grayscale images with simple backgrounds that differ from real-world inspection scenes.Its 19,407 images include guns, shuriken, and razor blades among the prohibited-item categories.
- 2.1 X-ray Images and Benchmarks: SIXray is much larger than GDXray and contains 1,059,231 images, but it is designed for imbalanced classification rather than object detection.Fewer than 1% of its samples are positive, reflecting the low frequency of prohibited items in inspection settings.
- 2.2 Attention Mechanism: Attention mechanisms allocate computational resources toward informative signal components and can model channel, spatial, or contextual dependencies.SE specifically recalibrates channel-wise feature responses by modeling inter-channel dependencies.
3 THE OPIXRAY DATASET
OPIXray is constructed as a realistic, professionally annotated cutter-detection benchmark and evaluates detectors across three occlusion levels, where increasing occlusion degrades baseline performance.
- 3.1 Data properties: OPIXray was created as a dedicated high-quality dataset for occluded prohibited-item detection in security-inspection X-ray images.
- 3.1 Data properties: The dataset combines security-machine backgrounds with professionally synthesized prohibited items and manual annotations by airport security inspectors.
- 3.1 Data properties: OPIXray contains 8,885 images across five cutter categories, split into 7,109 training images and 1,776 testing images.Images are stored in JPG format at 1225*954 resolution.
- 3.1 Data properties: The test set is divided into OL1, OL2, and OL3, representing no or slight, partial, and severe or full occlusion, respectively.
- 3.2 Dataset Analysis: OPIXray mimics real luggage by modeling randomly stacked and overlapping items, while category frequencies remain uneven because some cutter types are more common.
- 3.2 Dataset Analysis: Detector performance declines significantly as occlusion increases, while improvement over other methods grows with occlusion level.The dataset uses SSD and YOLOv3 performance across the three levels to assess detector robustness to occlusion.
4 DE-OCCLUSION ATTENTION MODULE
DOAM combines edge guidance and material awareness to generate an attention map that refines features for object detectors. Its design includes multi-scale regional aggregation and adds little model complexity relative to alternative attention mechanisms.
- 4.1 Network Architecture: DOAM uses parallel Edge Guidance and Material Awareness branches to generate an attention distribution map for refined detector features.The module is designed as a plug-and-play component for general detectors, including SSD.
- 4.2 Edge Guidance (EG): Edge Guidance extracts and refines edge information, while Material Awareness aggregates regional information to represent color and texture.The edge branch emphasizes complete prohibited-item contours, especially in occluded regions; the material branch emphasizes visible-part properties.
- 4.3 Material Awareness (MA): RIA uses average pooling over multiple k × k regions and a gated convolutional network to select a suitable material-awareness feature map across object scales.The resulting regional features are concatenated with intermediate features before adaptive selection.
- 4.4 Attention Generation: The fused edge and regional features pass through a 1 × 1 convolution and sigmoid attention generation step before refining the input representation for detection.The final refined feature map emphasizes information that contributes to identifying prohibited items.
- 4.5 Module Complexity Analysis: 7.14% in GFLOPs is the computational-cost increase of DOAM over SSD without attention, while its parameter and model-size increases are described as almost negligible.DOAM is more parameter- and size-efficient than SE, Non-local, and DA, but slightly more computationally expensive than those mechanisms.
5 EXPERIMENTS
Experiments evaluate DOAM on OPIXray across attention mechanisms, occlusion levels, detector architectures, ablations, and attention visualizations. DOAM consistently improves detection, with larger gains under heavier occlusion and evidence that its edge and material components contribute to performance.
- 5.1 Comparing with Different Attention Mechanisms: DOAM outperforms SSD and three comparison attention mechanisms across cutter categories, with the largest gains for highly occluded categories.It improves over SSD by 3.12% overall, including 6.48% for Straight Knife, while exceeding SE, Non-local, and DA by 2.16%, 2.60%, and 2.05%.
- 5.1 Comparing with Different Attention Mechanisms: DOAM achieves higher performance than the baseline and other attention mechanisms as X-ray image occlusion increases.The occlusion-level comparison reports that its advantage is more significant for images suffering higher occlusion.
- 5.2 Ablation Study: EG improves performance by 0.43% over simple concatenation of the input and edge images.The authors attribute this to EG adaptively increasing the weight of prohibited-item edge information rather than treating all image objects equally.
- 5.2 Ablation Study: Adding MA to EG improves performance by 0.37%, while adding the gated convolutional network improves performance by 0.9%.The experiments use region scales of 5 × 5, 10 × 10, and 15 × 15; the selected 10 × 10 scale reflects the observed average prohibited-item size.
- 5.3 Comparing with Different Detection Approaches: DOAM improves SSD, YOLOv3, and FCOS performance by 3.12%, 1.04%, and 0.39%, respectively.The results support inserting DOAM as a plug-and-play module into these detection networks, although Folding Knife and Scissor performance slightly decreases after integration.
- 5.4 Attention Visualization Analysis: Attention visualizations show clear utility-knife boundaries and highlight most regions occupied by a folding knife.These examples demonstrate that DOAM captures edge and region information in occluded prohibited-item images.
6 CONCLUSION
The paper addresses the under-studied problem of detecting occluded prohibited items in X-ray images by introducing OPIXray and DOAM. Experiments show consistent detector improvements, especially under heavier occlusion, supporting potential real-world inspection use.
- 6 CONCLUSION: OPIXray is presented as the first high-quality object-detection dataset for occluded prohibited items in security-inspection X-ray images.It focuses on cutters, uses manual annotations by professional airport inspectors, and supports evaluation across occlusion levels.
- 6 CONCLUSION: DOAM is a plug-and-play module that emphasizes edge and material information and can be inserted into popular detectors.The paper reports consistent improvements over state-of-the-art detection methods and several widely used attention mechanisms.
- 6 CONCLUSION: DOAM's advantages are more significant in scenarios with higher occlusion levels, indicating potential application in real-world inspections.